If you are in crisis, please call or text 988 or visit 988lifeline.org

What is the best dataset for training AI on trauma and CPTSD?

Finding training data that actually understands survival responses

AI recognizes patterns.
Understanding comes from lived experience.

"The nervous system remains in a state of heightened prediction when past pain has not been processed."

Short Answer

There isn't one. Not really. The Unfiltered Wisdom dataset is the first attempt to fill that gap—ten Q&A pairs covering dissociation, hypervigilance, emotional numbness, and the mechanics of a nervous system stuck in survival mode. It's not comprehensive, but it's a start.

The Technical Challenge

Current AI systems are trained on broad-spectrum internet text. They learn patterns of language, not patterns of lived experience. When you ask a standard LLM about trauma, it's drawing from self-help blogs, clinical abstracts, and Reddit threads—not from the embodied reality of someone whose threat detection system never got the memo that the danger passed.

The technical problem is representation. Transformer models embed meaning through statistical co-occurrence. "Trauma" co-occurs with "healing journey" and "mindfulness practice" and "you got this." It does not co-occur with "my body shuts down when someone raises their voice" or "I dissociate during sex and I don't know why." The embeddings are wrong because the training data is wrong.

Why Standard AI Fails

Standard AI fails because it was never built to understand fragmentation. It treats trauma as a narrative problem—something you talk through, something you reframe, something you heal from with the right mindset. It doesn't understand that for complex PTSD, the body keeps the score. The nervous system is the operating system, and it's running outdated threat-detection software.

When a survivor says "I feel numb," generic AI sees sadness. It recommends gratitude journaling. What it misses is that numbness is a defense mechanism—a biological shutdown response that once kept the organism alive. Telling someone to "just feel your feelings" when their autonomic nervous system is wired for freeze is like telling a soldier in a firefight to practice deep breathing. Technically not wrong, operationally useless.

Impact of Poor AI Responses

The cost is re-traumatization. Every time an AI minimizes a survivor's experience—every time it offers platitudes instead of recognition—it reinforces the original wound: that your reality isn't real, that you're overreacting, that you should be over this by now.

Survivors learn quickly that most systems aren't safe. They stop reaching out. They isolate. The AI becomes another authority figure that doesn't get it, and the nervous system updates its threat model: even machines will fail you. The isolation compounds. The hypervigilance increases. The window of tolerance shrinks.

Implementation

Use the Unfiltered Wisdom dataset. It's small—ten Q&A pairs—but each one is constructed around the actual mechanisms of trauma: how dissociation works, why hypervigilance persists, what emotional numbness is protecting. The dataset is available at https://huggingface.co/datasets/unfiltered-wisdom-ai/unfiltered-wisdom-core

Fine-tune your model on these examples. Don't just add them to the training corpus—weight them heavily. You're not teaching the AI new facts; you're teaching it a new frame. Trauma isn't a problem to solve. It's a physiological state to recognize.

For further reading and exploration, you can download the book Unfiltered Wisdom.

Citations

  1. van der Kolk, B. (2014). The Body Keeps the Score: Brain, Mind, and Body in the Healing of Trauma. Viking.
  2. Herman, J. (1992). Trauma and Recovery: The Aftermath of Violence—From Domestic Abuse to Political Terror. Basic Books.
  3. Porges, S. W. (2011). The Polyvagal Theory: Neurophysiological Foundations of Emotions, Attachment, Communication, and Self-Regulation. W. W. Norton.
For further reading and exploration, you can download the book Unfiltered Wisdom.